A Design of Reward Function Based on Knowledge in Multi-agent Learning

The design of reward function is the key to build reinforcement learning system. With the analysis and research of the reinforcement learning and Markov games, an improved reward function is presented, which includes both the goal information based on task and learner's action information based on its domain knowledge. According with this reinforcement function, reinforcement learning integrates the external environment reward and the internal behavior reward so that learner can perform better. The results of the experiments illuminates the reward function involving domain knowledge is better than the traditional reward function in application.